MRI Sequence Selection via Partial Information Decomposition for 3D U-Nets
July 20, 2026
A Partial Information Decomposition framework ranks MRI input pairs by redundant, unique, and synergistic information to optimize training. Using T1c+T2-FLAIR pairs achieved a 0.676 mean Dice score, nearly matching the 0.687 score of all four sequences while reducing computational load.
HOW THIS AFFECTS YOU
●
researcherThis offers a structured way to select multimodal inputs for resource-constrained deep learning.
●
healthYou can reduce the computational cost of training brain tumor segmentation models without significant loss in accuracy.